He was right a few years ago. But now there are several groups who have developed efficient capable online learning systems that don't require much data or iteration. When these and other existing types of cutting edge neural network advances such as techniques for avoiding catastrophic forgetting are combined with incremental training in diverse environments with general inputs and outputs, I believe we will see general purpose intelligence.
I believe we will see some demonstrations of AGI in the next two years. At first they will likely be general but unimpressive and not really as capable as animals or humans, and so people will dismiss them. But quickly the capabilities demonstrated will increase and before 2023-2024 there will likely be consensus that it has been achieved.
Look at systems like this one https://github.com/ogmacorp/EOgmaNeo. It's a whole other type of NN that Kasparov and others aren't even aware of.
Do you believe that artificial intelligence will be capable of deciding, given an algorithm and a set of inputs, whether the algorithm will finish running?
Here's a way of thinking about the undecidability of the halting problem. Let's say you've got a person who's amazing at reading minds, and you bring someone off the street and tell them they can either have steak or a cupcake (but not both). You then ask the mindreader to decide if the person will have the cupcake or the steak. Conceivably, they might be able to figure out which one the person will have. Now let's say that you add a twist: you walk up to the person from the street and tell them what the mindreader predicted; in that case, the mindreader can't succeed because the subject can choose to do the opposite. That's similar to how the undecidability proof of the halting problem works.
Now instead of a mindreader, we have a halting oracle, and to make its job impossible we have a test program that is "made aware" of the halting oracle, and does the opposite of what the oracle says. Impossible problem for the oracle. But that then begs the question, how many potential applications of the halting problem will involve test subjects that actually know what the halting oracle thinks? How many test subjects even know about the halting oracle? For instance, how can a program that looks for counterexamples to the Goldbach conjecture know anything about your halting oracle? In these cases, the undecidability proof doesn't apply.
So the answer to your question is conceivably yes.
Yes, that was Turing's proof. Church's more indirect proof is that it's impossible to prove the equivalence of two lambda expressions. But the real essence of the problem is more like, you can't generally know ahead of time all the values that will be presented inside loop bodies. If you try to actually elaborate the loop then you get caught again: if the elaboration of the loop keeps going on for a while, the problem is re-presented, at what point do you give up? Which is to say, will this program, with these inputs, run forever? But this is basically Church's proof, since this is also the question, am I in exactly the same configuration as before? Without an ability to decide lambda expression equivalence, that question is also hopeless.
Hence the work that gets done in this area constrains the problem down to situations in which you can know enough to decide halting, like traversal of lists or trees that are known to be finite. I bring this up because, when it comes to AI, people want more than this.
Yes, so given this, do you (i.e., people optimistic about AI) believe that computer programs will be able to generate meaningful, novel computer programs, given that even the most cursory subproblem is impossible? Obviously I don't just mean metaprogramming, but the sorts of things people want artificial intelligence to be able to do, the singularity and so forth.
Lol. Ok so you are taking this halting thing and think that it means generally that no computer program can predict what a computer program will do, and therefore that proves that we will never have computers writing programs, and therefore never have general intelligence. You are really misinterpreting that stuff and not thinking it through.
For starters take a look at the field of program synthesis. It's not AGI but it demonstrates the first thing you misunderstood.
It turns out that the problem of deciding whether a predicate is universally valid in an axiomatic system is also undecidable (and, appropriately enough, called the decision problem). That is to say, declarative systems "corresponding to" general computation are also undecidable. Which isn't really surprising, since logical recurrence is isomorphic to functional recursion. Hence also why the examples I could readily find of program synthesis are decidable problems, like deciding the maximum of two numbers or deciding membership in a list.
Back at the time when Deep Blue won chess match against Kasparov everyone in the media said about superior intelligence of Deep Blue.
While I at that time clearly realized that IBM just built brute-force "bulldozer" which can look for 200 million positions per second. Even with that power it had only a slight advantage over Kasparov who can look at only a handful of positions per second.
Now, we have another generation of "intelligent" machines based on deep learning but I see this as just upgraded version of brute-force "bulldozers". Now, it takes hundreds of millions of samples to infer the rules which human can infer from only a thousand or even less samples.
So I would call truly intelligent machine which can learn to play chess or go looking/playing only to a few thousands examples and calculating only a few moves ahead and not more than a few moves per second. Obviously that machine would beat human intelligence completely.
Although, such machine still may not have self-consciousness with qualia but this yet another big challenge.
Are you sure that human thought isn't basically brute-force bulldozing? Just because we don't feel like it is doesn't mean it is.
The time it takes us to learn something, the number of times we have to see/experience it could be akin to bulldozing couldn't it?
There are a lotttt of neurons in our brains that are constantly going off, perhaps comparable to the amount of transistors in a deep learning gpu if you account for the training time difference
Human chess players learn from each other. What might look at first like learning from a small sample is really a great deal of knowledge transferred via a small sample. Millions of people have played billions of chess games combined. We're learning by parallel Monte Carlo simulation.
One thing I'd be interested to learn is, how much of what makes the difference between an above average chess player and a Master or a Grandmaster can be tied to better decision making after looking 3 or 5 moves ahead, and how much is the Master/Grandmaster's ability to look 10+ moves ahead?
The looking 10 or even just 5 moves ahead thing is overstated and this is not actually how it works most of the time. Most GMs only calculate that far in the endgame. Before that, often looking 2 or 3 moves ahead is sufficient based on strategic elements or opening theory (which can't easily be understood by 'looking moves ahead'; they're things like, "this pawn is passed" or "my light squares will become very weak" which are can be substitutes for looking 30+ moves ahead).
Often positions resemble historic or previous games, so pattern recognition here and the themes (e.g., "this particular structure will make it easier to get my rook on the 7th rank at some point") of the old game are important.
In fact, Capablanca, a former World Champion and endgame expert has a famous quote claiming to only look 1 move ahead.
It has already been two decades. We’re suppose to be three decades from the singularity. Personally, it doesn’t feel like we’re accelerating towards an AI that surpasses humans, in general.
I believe we will see some demonstrations of AGI in the next two years. At first they will likely be general but unimpressive and not really as capable as animals or humans, and so people will dismiss them. But quickly the capabilities demonstrated will increase and before 2023-2024 there will likely be consensus that it has been achieved.
Look at systems like this one https://github.com/ogmacorp/EOgmaNeo. It's a whole other type of NN that Kasparov and others aren't even aware of.